Pith. sign in

REVIEW 4 major objections 5 minor 46 references

KARST: Multi-Kernel Kronecker Adaptation with Re-Scaling Transmission for Visual Classification

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read KARST claims a multi-kernel Kronecker adapter plus re-scaling transmission beats prior PEFT methods and full fine-tuning on VTAB-1K and few-shot benchmarks with negligible inference cost.

desk verdict Honest incremental PEFT paper with a correct multi-kernel Kronecker adapter; the missing Krona baseline and missing error bars weaken the central claim. read the letter →

arxiv 2502.06779 v1 pith:2FLVJ7GO submitted 2025-02-10 cs.CV cs.AI

Yue Zhu , Haiwen Diao , Shang Gao , Long Chen , Huchuan Lu This is my paper · ORCID
classification cs.CVcs.AI
keywords parameter-efficientfine-tuningKroneckerproductmulti-kerneladaptationre-scalingtransmissionvisualclassificationVTAB-1Kfew-shotlearningvisiontransformer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

KARST proposes a parameter-efficient fine-tuning recipe for pre-trained vision transformers: replace the weight update $\Delta W$ with a sum of $N$ Kronecker products of small trainable matrices, and route the merged output through learned channel-wise scaling and shift factors before it feeds the next frozen layer. The paper claims this combination gives the adapter access to several complementary low-rank subspaces rather than one, while the re-scaling factors repair the distribution mismatch that inserted adapters typically create with pre-trained features. On VTAB-1K the method reaches 78.1% with ViT-B/16 and 78.6% with Swin-B, using 0.33M and 0.45M trainable parameters, topping every compared PEFT method and full fine-tuning. Because every added term is linear, the whole module and the re-scaling can be folded back into the original weight matrices after training, so inference costs nothing extra. The broader claim is that these two ideas generalise: the re-scaling transmission alone also improves existing factorized adapters such as FacT.

What carries the argument

The load-bearing object is the multi-kernel Kronecker update $\Delta W = \sum_{i=1}^N C_i \otimes (A_i B_i)$, where $\otimes$ is the Kronecker (tensor) product that builds a large block-structured matrix from two smaller ones. The stacking dimension $m$ controls the size of $C_i$, and $A_i B_i$ is a low-rank factorization of the inner matrix, so the total parameter count stays small while the effective update spans several subspaces. This structure runs in parallel with the frozen pre-trained weight $W_0$. The second mechanism is re-scaling transmission, written $y = (s_1 + 1) \odot (W_0 + \Delta W)x + s_2$, with channel-wise scale and shift parameters initialized to zero so the pre-trained state is preserved at the start of training. Being entirely linear, the re-scaling can be absorbed into the preceding linear layer after training, which is why the method adds no inference cost.

What would settle it

Run KARST with $N=1$ versus $N=2$ at matched parameter counts on each of the 19 VTAB-1K tasks; if the single-kernel version matches or beats the multi-kernel version on most tasks, the multi-kernel component is not carrying the reported gain. A second check is to ablate the re-scaling transmission per task rather than only on the aggregate average.

Watch

Extended reading notes

Core claim

The central discovery is that making the adaptation update a sum of multiple Kronecker-product kernels, $\Delta W = \sum_{i=1}^N C_i \otimes (A_i B_i)$, with $C_i \in \mathbb{R}^{m \times m}$, $A_i \in \mathbb{R}^{(D_{in}/m) \times r}$, $B_i \in \mathbb{R}^{r \times (D_{out}/m)}$, and then applying channel-wise scale $(s_1 + 1)$ and shift $s_2$ to $(W_0 + \Delta W)x$, yields a parameter-efficient adapter that beats existing PETL methods and full fine-tuning on VTAB-1K (78.1% on ViT-B/16 and 78.6% on Swin-B with at most 0.45M trainable parameters) and on five fine-grained few-shot benchmarks. The multi-kernel sum separates the update into $N$ complementary compact subspaces, and the re-scaling factors align the merged features with the frozen pre-trained layers. All terms are linear, so after training the update and the re-scaling fold back into $W_0$, leaving inference unchanged.

Load-bearing premise

The single hyperparameter choice, $N=2$ kernels, stacking dimension $m=8$, and rank $r \le 8$, is selected from average VTAB-1K accuracy and then applied unchanged to all 19 tasks and all few-shot settings, while the paper reports no per-dataset or per-backbone sensitivity analysis.

Editorial extensions

If this is right

  • KARST matches or beats full fine-tuning with only 0.33M (ViT-B/16) or 0.45M (Swin-B) trainable parameters, so large-model adaptation no longer requires touching backbone weights.
  • Because all inserted terms are linear and re-parameterizable, deployed KARST models incur zero extra latency or memory at inference relative to the fine-tuned backbone.
  • The re-scaling transmission transfers: adding it to FacT raises VTAB-1K average accuracy by about 0.7 points, suggesting it is a standalone fix for adapter-to-backbone distribution misalignment.
  • Consistent wins on 1, 2, 4, 8, and 16-shot fine-grained benchmarks indicate that KARST is especially useful when labels are scarce.
  • The same update form works on both ViT and Swin transformer architectures, showing the mechanism is not tied to one backbone design.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The sum-of-kernels form is a cheap way to widen the effective spectrum of the adaptation without increasing the rank $r$; a natural testable extension is varying $N$ per layer or per dataset rather than using one global value.
  • The RST mechanism suggests that any linear PEFT that perturbs features, such as LoRA or Kronecker adapters, could gain from a learned rescale before the next frozen block; the paper only demonstrates this for FacT and its own KA module.
  • Since $s_1$ and $s_2$ fold into the previous linear layer, KARST effectively provides a structural way to update the affine statistics of a frozen network while also injecting low-rank subspaces, which could make it attractive for memory-constrained deployment.
  • The paper leaves open whether $N=2$ remains optimal at larger scales or with higher-rank budgets; the reported trend that accuracy rises with $N$ at increasing parameter cost suggests a budget-aware per-layer kernel count could improve the trade-off.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. This paper proposes KARST, a parameter-efficient fine-tuning method for vision transformers. KARST parameterizes the weight update as a sum of N Kronecker products, ΔW = Σ_i C_i ⊗ (A_i B_i), and adds channel-wise scale and shift factors (re-scaling transmission) to the output of the adapted linear layer. The authors report results on VTAB-1K with ViT-B/16 and Swin-B, and on five fine-grained few-shot datasets, claiming state-of-the-art accuracy with a negligible inference overhead due to re-parameterization into the pretrained weights.

Significance. If the results hold, KARST is a clean and practical contribution: the multi-kernel Kronecker sum is a simple extension of Krona with an explicit parameter-reduction factorization, the re-scaling transmission is trivially re-parameterizable, and the released code will facilitate reproduction. The mathematical formulation is straightforward and the re-parameterization argument is sound. However, the empirical support for the central claim is currently incomplete: the direct predecessor method is not compared, the key hyperparameter is tuned on the test benchmark, and the headline margins are below typical noise levels without error bars. These issues are addressable with additional experiments rather than conceptual flaws.

major comments (4)
  1. [III-B, Tables I-II] Krona [12], the single-kernel Kronecker adapter that this paper directly extends, is absent from all experimental comparisons. The multi-kernel mechanism is the central contribution (Sec. II-B), but Fig. 3 only varies the number of kernels inside the proposed C_i ⊗ (A_i B_i) parameterization; it does not compare against Krona's original Kronecker construction. Without this baseline, the reported 0.2 pp advantage over Sparse-Tuning cannot be attributed to the multi-kernel design rather than to the Kronecker low-rank structure or the re-scaling module, both of which have precedents (FacT, SSF). Please add Krona to Tables I and II, or otherwise ablate the kernel count against the original Krona formulation.
  2. [III-D, Fig. 3] The kernel number N is selected on the VTAB-1K benchmark itself ('After careful evaluation ... selected a kernel number of 2'), and this same configuration is then used for all 19 VTAB-1K tasks in Tables I and II. This is test-set-based hyperparameter selection and can inflate the reported averages. The choice should be made on held-out validation data (or per-dataset validation), and the sensitivity of the aggregate result to N (and to m and r) should be reported. As it stands, the main result depends on a single configuration whose selection procedure is not independent of the test set.
  3. [III-B, Tables I-II and Fig. 2] No variance estimates are reported for the VTAB-1K results, and the headline gain over Sparse-Tuning on ViT-B/16 is only 0.2 pp (78.1 vs 77.9). Without multiple seeds, this difference is within the typical run-to-run noise for this benchmark, so the claim that KARST 'outperforms' state-of-the-art PEFT methods is not yet established at this margin. Please report mean ± std over at least three seeds for the main comparisons, and provide numerical per-dataset results for the few-shot benchmark instead of only a figure.
  4. [II-B, Eq. (3) and Table III] The low-rank factorization D_i = A_i B_i in Eq. (3) is introduced as a parameter-reduction step, but the paper never measures its effect on accuracy. Table III ablates the re-scaling module only. Since Eq. (2) with full D_i is more expressive, the paper should include an ablation comparing Eq. (2) with Eq. (3) (at a matched parameter budget or with the parameter count stated) to verify that the factorization does not degrade accuracy beyond the parameter savings. This is directly relevant to isolating the source of the gains.
minor comments (5)
  1. [Introduction, first paragraph] 'Inspire by this' should be 'Inspired by this'.
  2. [II-C, Eq. (4)] The symbol '1' in Eq. (4) is not defined; specify that it is a vector of ones of the appropriate dimension.
  3. [II-B, after Eq. (2)] The claim that 'the resulting ΔW is not rank deficient' is imprecise because a sum of Kronecker products can be rank deficient; please state the intended condition or justify it from the initialization and training.
  4. [Tables I and II, headers] The two average columns at the right are not clearly labeled; please define 'Average' and 'All Set Average' in the caption or table header so the reader knows which groups they cover.
  5. [Fig. 2] The few-shot curves are difficult to read in the provided figure; include numeric values or a supplementary table with the per-dataset results and error bars.

Circularity Check

1 steps flagged · score 4.0 of 10

The method itself is not derived circularly, but one headline number is partially circular: the kernel number N=2 is selected on the VTAB-1K average accuracy and that same average is then reported as the main result.

  1. fitted input called prediction [Sec. III-D (Fig. 3), with hyperparameters stated in Sec. III-B/III-C and results in Table I]
    "Fig. 3 demonstrate the influence of different kernel numbers for multi-kernel Kronecker Adaptation. The ViT-B/16 is adopted as the baseline on the VTAB-1K benchmark. We observe that as the number of kernels increases, the average accuracy improves. However, this also results in a corresponding rise in the number of fine-tuned parameters. After careful evaluation of the trade-off between performance and parameter efficiency, we selected a kernel number of 2 for our experiments."

    Figure 3 chooses the kernel number N=2 by looking at the VTAB-1K average accuracy, and Table I then reports KARST's VTAB-1K average accuracy (78.1%) as the headline evidence that it outperforms Sparse-Tuning and DTL. The hyperparameter is therefore fitted to the exact evaluation metric that is later presented as the prediction. This makes the reported VTAB-1K gain partially a selection artifact rather than an out-of-sample result. The few-shot experiments use N=2 without retuning on those datasets, so they provide a weaker but not circular transfer check. This is evaluation leakage on one hyperparameter, not a derivation identity, so the circularity is moderate rather than total.

full rationale

The paper's central equations (2)-(4) define a parameterization and are not derived from the results they are used to explain. The self-citations in the introduction and the citation of Krona [12] are contextual and not load-bearing; no uniqueness theorem or ansatz is imported from the authors' prior work to force the design. The absence of a direct Krona baseline is an experimental omission, not circularity. The one genuine circularity-adjacent issue is the selection of N=2 using the VTAB-1K average accuracy, followed by reporting that same benchmark's average as the headline result. Because this affects only one hyperparameter and the few-shot benchmark provides some independent support, the overall circularity score is low but not zero.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical or conceptual entities are introduced. The method is a composition of existing linear algebra constructs. The main uncharged assumptions are the distribution-alignment benefit of the re-scaling mechanism and the generality of the chosen hyperparameters across all datasets and backbones.

free parameters (3)
  • kernel number N = 2
    Selected based on VTAB-1K average accuracy vs. parameter trade-off in Fig. 3, then used for all experiments. The choice is a hyperparameter tuned on the evaluation benchmark.
  • stacking dimension m = 8
    Set to 8 and used throughout. No sensitivity analysis for m is reported, so the value is a fixed design choice.
  • rank r = ≤8
    The rank bound for the low-rank factors A_i and B_i is stated as r≤8 for both LoRA and KARST. Exact per-layer ranks are not specified.
assumptions (3)
  • standard math Kronecker product algebra and standard matrix rank decompositions are valid for representing ΔW.
    Used in Eq. 1-3 as the core representation of the weight update.
  • domain assumption Pre-trained ViT and Swin features are a suitable starting point for all 19 VTAB-1K and 5 few-shot tasks.
    The evaluation protocol assumes that fine-tuning a frozen ImageNet-21k backbone is the right setting; this is standard in PEFT but never justified in the paper.
  • ad hoc to paper Adding channel-wise affine scaling to the merged output can realign intermediate feature distributions with subsequent pre-trained layers.
    Section II-C asserts this mechanism, but the paper provides no measurement of distribution shift and no theoretical support. The evidence is limited to an ablation showing a 0.5 to 0.7 point gain.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KARST: Multi-Kernel Kronecker Adaptation with Re-Scaling Transmission for Visual Classification." pith.science (2026). https://pith.science/paper/2FLVJ7GO

@misc{pith2026250206779,
  author       = {Pith},
  title        = {Pith review of: KARST: Multi-Kernel Kronecker Adaptation with Re-Scaling Transmission for Visual Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2FLVJ7GO}},
  note         = {Machine review of arXiv:2502.06779}
}
read the original abstract

Fine-tuning pre-trained vision models for specific tasks is a common practice in computer vision. However, this process becomes more expensive as models grow larger. Recently, parameter-efficient fine-tuning (PEFT) methods have emerged as a popular solution to improve training efficiency and reduce storage needs by tuning additional low-rank modules within pre-trained backbones. Despite their advantages, they struggle with limited representation capabilities and misalignment with pre-trained intermediate features. To address these issues, we introduce an innovative Multi-Kernel Kronecker Adaptation with Re-Scaling Transmission (KARST) for various recognition tasks. Specifically, its multi-kernel design extends Kronecker projections horizontally and separates adaptation matrices into multiple complementary spaces, reducing parameter dependency and creating more compact subspaces. Besides, it incorporates extra learnable re-scaling factors to better align with pre-trained feature distributions, allowing for more flexible and balanced feature aggregation. Extensive experiments validate that our KARST outperforms other PEFT counterparts with a negligible inference cost due to its re-parameterization characteristics. Code is publicly available at: https://github.com/Lucenova/KARST.

Figures

Figures reproduced from arXiv: 2502.06779 by the authors.

Figure 1
Figure 1. Framework of our proposed KARST. We first transform hidden [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Top-1 accuracy on fine-grained few-shot benchmark with ViT-B/16 as the backbone. Note that our KARST significantly outperforms other PETL [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Average accuracy of KARST on VTAB-1K with multiple kernels. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 36 canonical work pages

  1. [12]

    Krona: Parameter efficient tuning with kronecker adapter,

    A. Edalati, M. Tahaei, I. Kobyzev, V . P. Nia, J. J. Clark, and M. Reza- gholizadeh, “Krona: Parameter efficient tuning with kronecker adapter,” arXiv preprint arXiv:2212.10650 , 2022. 1, 2

    arXiv 2022

  2. [1]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” in ICLR, 2021. 1, 3

    work page 2021

  3. [2]

    Unveiling encoder-free vision-language models,

    H. Diao, Y . Cui, X. Li, Y . Wang, H. Lu, and X. Wang, “Unveiling encoder-free vision-language models,” arXiv:2406.11832, 2024. 1

    arXiv 2024

  4. [3]

    Similarity reasoning and filtration for image-text matching,

    H. Diao, Y . Zhang, L. Ma, and H. Lu, “Similarity reasoning and filtration for image-text matching,” in AAAI, 2021, pp. 1218–1226. 1

    work page 2021

  5. [4]

    Unipt: Universal parallel tuning for transfer learning with efficient parameter and memory,

    H. Diao, B. Wan, Y . Zhang, X. Jia, H. Lu, and L. Chen, “Unipt: Universal parallel tuning for transfer learning with efficient parameter and memory,” in CVPR, 2024. 1

    work page 2024

  6. [5]

    Sherl: Synthesizing high accuracy and efficient memory for resource- limited transfer learning,

    H. Diao, B. Wan, X. Jia, Y . Zhuge, Y . Zhang, H. Lu, and L. Chen, “Sherl: Synthesizing high accuracy and efficient memory for resource- limited transfer learning,” in ECCV, 2025, pp. 75–95. 1

    work page 2025

  7. [6]

    Bitfit: Simple parameter- efficient fine-tuning for transformer-based masked language-models,

    E. B. Zaken, Y . Goldberg, and S. Ravfogel, “Bitfit: Simple parameter- efficient fine-tuning for transformer-based masked language-models,” in ACL, 2022, pp. 1–9. 1, 3

    work page 2022

  8. [7]

    Visual prompt tuning,

    M. Jia, L. Tang, B. Chen, C. Cardie, S. J. Belongie, B. Hariharan, and S. Lim, “Visual prompt tuning,” in ECCV, vol. 13693, 2022, pp. 709–

    work page 2022

Show all 46 references
  1. [8]

    Gssf: Generalized structural sparse function for deep cross-modal metric learning,

    H. Diao, Y . Zhang, S. Gao, J. Zhu, L. Chen, and H. Lu, “Gssf: Generalized structural sparse function for deep cross-modal metric learning,” IEEE TIP, 2024. 1

    2024

  2. [9]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” in ICLR, 2022. 1, 2, 3

    2022

  3. [10]

    Fact: Factor-tuning for lightweight adaptation on vision transformer,

    S. Jie and Z. Deng, “Fact: Factor-tuning for lightweight adaptation on vision transformer,” in AAAI, 2023. 1, 2, 3

    2023

  4. [11]

    Compacter: Efficient low- rank hypercomplex adapter layers,

    R. K. Mahabadi, J. Henderson, and S. Ruder, “Compacter: Efficient low- rank hypercomplex adapter layers,” in NeurIPS, 2021, pp. 1022–1035. 1

    2021

  5. [13]

    Superlora: Parameter-efficient unified adaptation of multi-layer attention modules,

    X. Chen, J. Liu, Y . Wang, M. Brand, G. Wang, T. Koike-Akino et al. , “Superlora: Parameter-efficient unified adaptation of multi-layer attention modules,” arXiv preprint arXiv:2403.11887 , 2024. 1

    2024 arXiv

  6. [14]

    One-for-all: Generalized lora for parameter-efficient fine-tuning,

    A. Chavan, Z. Liu, D. Gupta, E. Xing, and Z. Shen, “One-for-all: Generalized lora for parameter-efficient fine-tuning,” arXiv preprint arXiv:2306.07967, 2023. 1, 3

    2023 arXiv

  7. [15]

    Multilora: Democratizing lora for better multi-task learning,

    Y . Wang, Y . Lin, X. Zeng, and G. Zhang, “Multilora: Democratizing lora for better multi-task learning,” arXiv preprint arXiv:2311.11501 , 2023. 1

    2023 arXiv

  8. [16]

    Adaptive budget allocation for parameter-efficient fine-tuning,

    Q. Zhang, M. Chen, A. Bukharin, P. He, Y . Cheng, W. Chen, and T. Zhao, “Adaptive budget allocation for parameter-efficient fine-tuning,” in ICLR, 2023. 2

    2023

  9. [17]

    A large-scale study of representation learning with the visual task adaptation benchmark,

    X. Zhai, J. Puigcerver, A. Kolesnikov, P. Ruyssen, C. Riquelme, M. Lu- cic, J. Djolonga, A. S. Pinto, M. Neumann, A. Dosovitskiy et al. , “A large-scale study of representation learning with the visual task adaptation benchmark,” arXiv preprint arXiv:1910.04867 , 2019. 3

    1910 arXiv

  10. [18]

    One-shot learning of object cate- gories,

    L. Fei-Fei, R. Fergus, and P. Perona, “One-shot learning of object cate- gories,” IEEE transactions on pattern analysis and machine intelligence, vol. 28, no. 4, pp. 594–611, 2006. 3

    2006

  11. [19]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hinton et al. , “Learning multiple layers of features from tiny images,” 2009. 3

    2009

  12. [20]

    Describing textures in the wild,

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi, “Describing textures in the wild,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2014, pp. 3606–3613. 3

    2014

  13. [21]

    Automated flower classification over a large number of classes,

    M.-E. Nilsback and A. Zisserman, “Automated flower classification over a large number of classes,” in 2008 Sixth Indian conference on computer vision, graphics & image processing . IEEE, 2008, pp. 722–729. 3

    2008

  14. [22]

    Cats and dogs,

    O. M. Parkhi, A. Vedaldi, A. Zisserman, and C. Jawahar, “Cats and dogs,” in 2012 IEEE conference on computer vision and pattern recog- nition. IEEE, 2012, pp. 3498–3505. 3

    2012

  15. [23]

    Sun database: Large-scale scene recognition from abbey to zoo,

    J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba, “Sun database: Large-scale scene recognition from abbey to zoo,” in 2010 IEEE computer society conference on computer vision and pattern recognition. IEEE, 2010, pp. 3485–3492. 3

    2010

  16. [24]

    Reading digits in natural images with unsupervised feature learning,

    Y . Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, A. Y . Ng et al. , “Reading digits in natural images with unsupervised feature learning,” in NIPS workshop on deep learning and unsupervised feature learning , vol. 2011, no. 2. Granada, 2011, p. 4. 3

    2011

  17. [25]

    Remote sensing image scene classifi- cation: Benchmark and state of the art,

    G. Cheng, J. Han, and X. Lu, “Remote sensing image scene classifi- cation: Benchmark and state of the art,” Proceedings of the IEEE , vol. 105, no. 10, pp. 1865–1883, 2017. 3

    2017

  18. [26]

    Eurosat: A novel dataset and deep learning benchmark for land use and land cover classi- fication,

    P. Helber, B. Bischke, A. Dengel, and D. Borth, “Eurosat: A novel dataset and deep learning benchmark for land use and land cover classi- fication,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 12, no. 7, pp. 2217–2226, 2019. 3

    2019

  19. [27]

    Rotation equivariant cnns for digital pathology,

    B. S. Veeling, J. Linmans, J. Winkens, T. Cohen, and M. Welling, “Rotation equivariant cnns for digital pathology,” in Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Par...

    2018

  20. [28]

    Kaggle diabetic retinopathy detection,

    E. Kaggle, “Kaggle diabetic retinopathy detection,” July 2015, accessed: Sep. 12, 2024. [Online]. Available: https://www.kaggle.com/ c/diabetic-retinopathy-detection 3

    2015

  21. [29]

    Clevr: A diagnostic dataset for compositional language and elementary visual reasoning,

    J. Johnson, B. Hariharan, L. Van Der Maaten, L. Fei-Fei, C. Lawrence Zitnick, and R. Girshick, “Clevr: A diagnostic dataset for compositional language and elementary visual reasoning,” in Proceed- ings of the IEEE conference on computer vision and pattern recognition, 2017, pp...

    2017

  22. [30]

    dsprites: Disen- tanglement testing sprites dataset,

    L. Matthey, I. Higgins, D. Hassabis, and A. Lerchner, “dsprites: Disen- tanglement testing sprites dataset,” 2017. 3

    2017

  23. [31]

    Learning methods for generic object recognition with invariance to pose and lighting,

    Y . LeCun, F. J. Huang, and L. Bottou, “Learning methods for generic object recognition with invariance to pose and lighting,” in Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. , vol. 2. IEEE, 2004, pp. II–104. 3

    2004

  24. [32]

    Deepmind lab,

    C. Beattie, J. Z. Leibo, D. Teplyashin, T. Ward, M. Wainwright, H. K¨uttler, A. Lefrancq, S. Green, V . Vald´es, A. Sadik et al., “Deepmind lab,” arXiv preprint arXiv:1612.03801 , 2016. 3

    2016 arXiv

  25. [33]

    Vision meets robotics: The kitti dataset. the international journal of robotics research,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The kitti dataset. the international journal of robotics research,” Int. J. Rob. Res, pp. 1–6. 3

  26. [34]

    Fine- grained visual classification of aircraft,

    S. Maji, E. Rahtu, J. Kannala, M. Blaschko, and A. Vedaldi, “Fine- grained visual classification of aircraft,” arXiv preprint arXiv:1306.5151,

    arXiv

  27. [35]

    Food-101–mining discriminative components with random forests,

    L. Bossard, M. Guillaumin, and L. Van Gool, “Food-101–mining discriminative components with random forests,” in Computer Vision– ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13 . Springer, 2014, pp. 446–461. 3

    2014

  28. [36]

    3d object representations for fine-grained categorization,

    J. Krause, M. Stark, J. Deng, and L. Fei-Fei, “3d object representations for fine-grained categorization,” in Proceedings of the IEEE interna- tional conference on computer vision workshops , 2013, pp. 554–561. 3

    2013

  29. [37]

    A visual vocabulary for flower classification,

    M.-E. Nilsback and A. Zisserman, “A visual vocabulary for flower classification,” in 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06), vol. 2, 2006, pp. 1447–1454. 3

    2006

  30. [38]

    Neural prompt search,

    Y . Zhang, K. Zhou, and Z. Liu, “Neural prompt search,” arXiv: 2206.04673, 2022. 3

    2022 arXiv

  31. [39]

    Do imagenet clas- sifiers generalize to imagenet?

    B. Recht, R. Roelofs, L. Schmidt, and V . Shankar, “Do imagenet clas- sifiers generalize to imagenet?” in International conference on machine learning. PMLR, 2019, pp. 5389–5400. 3

    2019

  32. [40]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in ICCV. IEEE, 2021, pp. 9992–10 002. 3

    2021

  33. [41]

    LST: ladder side-tuning for parameter and memory efficient transfer learning,

    Y . Sung, J. Cho, and M. Bansal, “LST: ladder side-tuning for parameter and memory efficient transfer learning,” in NeurIPS, 2022. 3

    2022

  34. [42]

    Adapt- former: Adapting vision transformers for scalable visual recognition,

    S. Chen, C. Ge, Z. Tong, J. Wang, Y . Song, J. Wang, and P. Luo, “Adapt- former: Adapting vision transformers for scalable visual recognition,” in NeurIPS, 2022. 3

    2022

  35. [43]

    Scaling & shifting your features: A new baseline for efficient model tuning,

    D. Lian, D. Zhou, J. Feng, and X. Wang, “Scaling & shifting your features: A new baseline for efficient model tuning,” in NeurIPS, 2022. 3

    2022

  36. [44]

    Dtl: Disentangled transfer learning for visual recognition,

    M. Fu, K. Zhu, and J. Wu, “Dtl: Disentangled transfer learning for visual recognition,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 11, 2024, pp. 12 082–12 090. 3

    2024

  37. [45]

    Heat: Head-level parameter efficient adaptation of vision transformers with taylor-expansion importance scores,

    Y . Zhong and Y . Zhou, “Heat: Head-level parameter efficient adaptation of vision transformers with taylor-expansion importance scores,” arXiv preprint arXiv:2404.08894, 2024. 3

    2024 arXiv

  38. [46]

    Sparse- tuning: Adapting vision transformers with efficient fine-tuning and inference,

    T. Liu, X. Liu, L. Shi, Z. Xu, S. Huang, Y . Xin, and Q. Yin, “Sparse- tuning: Adapting vision transformers with efficient fine-tuning and inference,” arXiv preprint arXiv:2405.14700 , 2024. 3

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.